glitch003

Author here. The short version: Qwen3.8's 248k-token vocabulary contains scraped-junk tokens that got almost no training (embedding norms down to 0.001 against a vocabulary mean of 0.899). Put one in a slot that matters, like a username, order ref, or SKU, and the model swaps in a hallucination while keeping the output schema-valid. An order ref of "Kinhted" came back as "order_id": "shelled", 16 out of 16 times at temperature 0.7.

We ran six control models. Mistral and Llama have the same class of problem and lose 5 to 7 of their 12 lowest-norm tokens. Both Gemma generations are clean, so a curated vocabulary avoids this entirely.

Happy to answer methodology questions. If you want to know whether a specific checkpoint has this problem, name it and I'll run the scan.

she called her husband, from jail, and asked him to move them so they couldn't be used in court against her. she said "do whatever you have to do". the cops listened to the phone call, watched him load them into the car and arrested him on the drive. i think it's pretty straightforward that she was asking for his help in obstructing the investigation by hiding potential evidence against her.

How does this compare to libp2p?

fwiw, tesla, the biggest "no physical buttons" company, does not have this problem at all. it's always very responsive, window buttons always work, there's no "startup" delay at all.

so i guess the choice is between companies that do tech poorly (legacy automakers) and one that does it right but with a big touch screen.

it's still USDS, they just changed the "D" from digital to DOGE

we get the whole US AID budget back, aside from PEPFAR. it's a good example of a case where they cut too much and undid some cuts.

surprisingly, it works excellent in rain. i was on the highway in florida in a torrential downpour, the kind where everyone turns their flashers on and goes 25mph. many people pulled over to wait it out. FSD handled it perfectly even though i could barely see. it did show a warning that it might be degraded due to weather, but there was no change in performance.

i can't speak for snow tho.

Hi HN - we’re building Payout.Team for freelancers, consultants, and contractors who work on projects together.

Here’s how it works: a users sends an invoice to their client, when the client pays, the funds are automatically routed to the members of the team / subcontractors. Payout.Team takes care of all the on-boarding and tax paper work for the subcontractors, including internationally.